A lightweight WebAssembly runtime that is fast, secure, and standards-compliant
This example project demonstrates using the wasi-nn API to perform PyTorch based inference. It consists of Rust code that is built using the wasm32-wasip1 target.
To run this example:
- Ensure you set appropriate Libtorch environment variables according to tch-rs instructions.
- Requires the C++ PyTorch library (libtorch) in version v2.4.0 to be available on your system.
export LIBTORCH=/path/to/libtorch
- Build Wasmtime with
wasmtime-wasi-nn/pytorchfeature. - Navigate to this example directory
crates/wasi-nn/examples/classification-example-pytorch. - Download
squeezenet1_1.ptmodel
curl https://github.com/rahulchaphalkar/libtorch-models/releases/download/v0.1/squeezenet1_1.pt --output fixture/model.pt -L
- Build this example
cargo build --target=wasm32-wasip1. - Run the generated wasm file with wasmtime after mapping the directory containing squeezenet1.1
model.ptand sample imagekitten.png${Wasmtime_root_dir}/target/debug/wasmtime -S nn --dir ${Wasmtime_root_dir}/crates/wasi-nn/examples/classification-example-pytorch::. ${Wasmtime_root_dir}/crates/wasi-nn/examples/classification-example-pytorch/target/wasm32-wasip1/debug/wasi-nn-example-pytorch.wasm - Check that result
281has highest probability, which corresponds totabby cat.